Enhancing financial inclusion measurements: A comparative analysis using unsupervised machine learning methods across Mexican municipalities
Academic Article in Scopus
While several different statistical methods have been employed to measure financial inclusion, two of the most frequently used in the literature are the adaptation of the Human Development Index (HDI) methodology proposed by Sarma (2008) and, more recently, the two-stage Principal Component Analysis (PCA) suggested by Cámara and Tuesta (2014). Employing a unique dataset on the access and usage of financial services across Mexican municipalities, this paper proposes the alternative use of two unsupervised Machine Learning (ML) methods: Nonnegative Matrix Factorization (NMF) and Kernel PCA (KPCA). To contrast the performance of the four different methods, we first cluster the municipalities using the K-Means algorithm. Thereafter, we focus our comparison of the indices produced by the four methods on the most financially excluded clusters, using standard ML performance metrics: precision, recall, F1 score, and accuracy. Our results show that both HDI and NMF significantly outperform PCA-based methods across different thresholds at the bottom of the distribution. Several robustness checks confirm our findings. This result is particularly significant, as poverty alleviation, one of the primary targets of financial inclusion, can only be achieved successfully if policymakers accurately identify individuals and regions that are excluded or underserved by the financial sector.
